What Anthropic’s Agent Usage Data Reveals

Edition 160 | February 23, 2026

Agent autonomy is like figure skating: grace comes from training, not improvisation.

Welcome back to Building AI Agents, your biweekly guide to what’s new in agentic AI!

In today’s issue…

• What Anthropic learned from millions of real agent interactions

• Why Gemini 3.1 Pro’s reasoning upgrade matters for agent reliability

• How an Amazon agent took down production, and the guardrails that failed

• Turning a Raspberry Pi into a local AI agent with OpenClaw

• Stripe’s playbook for running agents at scale

…and more

🔍 SPOTLIGHT

Here's a finding that caught me off guard. Anthropic just published a research paper analyzing millions of real human-agent interactions across Claude Code and their public API. The goal was to answer a deceptively simple question:

How much autonomy do people actually give their agents? The answer definitely challenges some assumptions.

The headline stat: experienced users auto-approve agent actions at more than double the rate of new users, over 40% of sessions versus roughly 20%. That's not surprising. People build trust over time. What is surprising is that those same experienced users also interrupt their agents nearly twice as often. New users interrupt about 5% of turns. Experienced users, around 9%.

That seems contradictory until you think about what it actually means. New users babysit. They approve every single action before it happens, which means they rarely need to step in mid-task. Experienced users flip the model entirely. They let the agent run, monitor what it's doing, and course-correct when something looks off. It's like the difference between micromanaging an intern and managing a capable employee, you give them more room, but you also develop sharper instincts for when to step in.

The research also uncovered what Anthropic calls a "deployment overhang": a gap between what models can handle and what people actually let them do. Among the longest-running sessions, the time agents worked autonomously before stopping nearly doubled in three months, from under 25 minutes to over 45 minutes. And that increase was smooth across model releases, which means it wasn't driven by smarter models. It was driven by users gradually expanding the boundaries of what they trusted agents to do. The implication: most people are significantly underusing their agents.

But the part of the paper that should matter most to you is where agents are actually being deployed, and where they aren't yet. Software engineering accounts for nearly 50% of all agentic activity on Anthropic's API. After that, the drop-off is steep. Business intelligence, customer service, sales, finance, and e-commerce each make up only a few percentage points of traffic. The full distribution looks like this:

That chart is a snapshot of where we are right now. Developers were first because code is easy to test, you run it and see if it works. But agents are starting to creep into domains where verifying output is harder and the stakes are higher. Healthcare. Finance. Cybersecurity. Anthropic found that 80% of tool calls had some form of safeguard in place and only 0.8% of actions appeared irreversible, reassuring numbers for now. But as adoption spreads beyond software, those numbers will shift, and the margin for error shrinks.

If you're looking for where to build or deploy agents next, that chart is your map. The industries with the smallest bars aren't empty because agents can't work there. They're small because nobody's built the right systems around them yet. The models are ready. The harnesses (the guardrails, the verification loops, the oversight workflows) are what's lagging behind.

As always, keep learning and building!

—AP